AdaptiveInvolutionNet: Spatially-adaptive involution with channel-wise attention for breast MRI tumor classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Accurate breast tumor classification from MRI is critical for patient outcomes.
• Conventional deep learning models struggle with subtle, spatially variant features in medical images.

Purpose of the Study:

• Introduce AdaptiveInvolutionNet (AIN), a novel deep learning framework.
• Enhance discriminative feature learning for breast tumor classification using hybrid architecture.

Main Methods:

• Utilize spatially-adaptive involution kernels for fine-grained feature extraction in early layers.
• Integrate channel-wise attention (squeeze-and-excitation modules) to amplify relevant features.
• Train and evaluate on 6,000 breast MRI images using five-fold cross-validation.

Main Results:

• Achieve a high test accuracy of 97% with consistent performance across cross-validation folds (average 96% ± 1%).
• Demonstrate strong agreement with true labels (Cohen's Kappa 0.93 ± 0.01).
• Produce trustworthy predictions with a low Brier score of 0.0241.

Conclusions:

• AIN successfully combines adaptive spatial feature extraction and attention mechanisms for advanced medical image analysis.
• The model shows high accuracy, robust generalization, and consistent reliability.
• AIN holds significant potential as a dependable computer-aided diagnostic tool for clinical breast cancer detection.